Llama (4 Scout / Maverick)

What it is

Llama is Meta's family of open-weight language models, one of the most widely self-hosted and fine-tuned model lines in the industry. Because the weights are downloadable (see Open Weight vs Closed Weight Models), Llama underpins a huge number of third-party products, research projects, and fine-tuned variants that never touch Meta's own infrastructure. The current generation, Llama 4, ships in two released sizes — Scout and Maverick — both using a Mixture-of-Experts design; a larger "Behemoth" variant was previewed but has not been publicly released.

A strategic shift at Meta

In April 2026, Meta's newly formed Superintelligence Labs shipped Muse Spark (see its own entry) — Meta's first proprietary, closed-weight flagship model. This marks a notable departure from Meta's "open science" positioning: Llama remains open weight and available today, but it's no longer confirmed to be the direction of Meta's most capable future models. Watch this entry and Muse Spark's for how that strategy develops.

Where these numbers come from

The benchmarks above are for Llama 4 Maverick, released April 2025. This entry should be updated via the admin dashboard if Meta ships a further open-weight Llama generation, or if the balance between the open Llama line and the closed Muse Spark line shifts further.

Notable traits

• Open weights with a permissive-but-not-unrestricted license (very large commercial deployments have historically required a separate agreement with Meta).
• Ships in multiple sizes, letting users trade capability for the ability to run smaller versions on modest hardware.
• One of the most fine-tuned base models in the open-source ecosystem — countless specialized variants exist built on top of various Llama generations.